fast 35b STRIP-OFF isolation arm — 7/9 solved, context strip confirmed as the regressor
Isolation rerun of the tangled 2026-07-08 night batch: SubagentsPi with
loop_guard=on subagent_nudges=on kept, STRIP_THINKING=off
KEEP_TOOL_RESULTS=0 (the only change vs fast__…20260708-034239).
Result: 7/9 solved — 6 PASS + query-optimize verifier reward 1.0 hidden
behind an ERR AgentTimeoutError — vs 3/9 with the strip on and 7/9 on the
07-07 baseline. Full recovery: fix-git, cancel-async-tasks,
sparql-university, query-optimize all came back. The context strip is the
confirmed 35b score regressor; loop guard + nudges are score-neutral.
Odd ones out: filter-js-from-html (fails every 35b run, 28s instant fail
here) and sanitize-git-repo (0/2 across both 07-08 arms after 2/2 pre-07-08
— possible flake or a real guard/nudge casualty, watch next batch).
query-optimize's ERR is two harness holes, not a model failure (analysis
updated in analysis/qwen3.6-35b-a3b/pi-subagents/query-optimize.md):
a planner child looped 17 min until the loop guard exit-1'd it (3rd
occurrence of that shape, and in a strip-OFF arm — orthogonal to the strip),
then after a clean ~4-min solve and final answer the pi process hung ~8 min
to the 30m cap (2nd confirmed post-session hang; both after subagent use).
Passing-trial times are otherwise healthy (sparql 1m52s, nginx 2m00s,
fix-git 2m53s, openssl 3m10s, regex 6m14s, cancel-async 6m30s agent).